Unclassified
31 · Testing & Verification
How the platform's execution, reconciliation, agent, memory, and projection behavior is tested — unit, contract, integration, runtime, end-to-end.
What Is Tested
| Level | Covers |
|---|---|
| Unit | Canonical ontology, normalizer, authority resolver, Work state machine |
| Contract | Loader/normalizer/pipeline/governed-writer, capability, projection, governance verdict shapes |
| Integration | Connector → pipeline → Spine, Capability Fabric → Governance → Hermes, re-ingest path |
| Runtime | Wiring (live vs orphaned — G-07), Model routing / ADK, MCP Pool routing boundary |
| Execution | Plan → approval → Hermes → evidence for happy, denied, and failure paths |
| Reconciliation | Expected vs observed mismatch, partial success, stale-read retries, compensation |
| Agent | Twin and Spine Agent Work ownership, handoff, evidence, continuity |
| Memory | Promotion boundaries (personal/AI → organizational), G-10 substrate behavior |
| Projection | Active Context Bundle assembly, scoping, invalidation, rehydration |
| End-to-end | Business Intent → Context → Work → Govern → Execute → Verify → Spine → Rehydrate under real tenant isolation |
Verification and reconciliation are logic, not optimism. They are tested against mismatches and partial successes, not just happy paths.
Where to Go Next
<CardGroup cols={2}>Why provider success alone is not the E2E assertion.
What every test asserts against.
How production confirms what tests predicted.